Single and multiphase flow leak detection in onshore/offshore pipelines and subsurface sequestration sites: An overview

Leaks may occur in existing pipelines, even when designed with quality construction and appropriate regulations. The economic impact of oil spills and natural gas dispersion from leaks can be huge. Failure to detect pipeline leaks promptly will have an adverse impact on life, the economy, the environment


Introduction
Despite having been built with quality materials and following regulations, existing pipelines can still develop leaks.Oil spills and the leakage of natural gas can have significant negative economic effects.Rapid pipeline leak detection is essential to prevent serious consequences for human health, the economy, the environment, and brand reputation.Therefore, for effective hydrocarbon transportation through a pipeline, both in onshore and offshore applications, early detection of leaks, their location, and their size with a high degree of sensitivity and reliability is crucial.Although some studies provide a summary of the different leak detection techniques for single-phase flow (Adnan et al., 2015;Zaman et al., 2020).However, there is a gap in the literature that studies and summarizes the different multiphase leak experiments, computational fluid dynamics (CFD) simulation, mechanistic modeling, machine learning, and digital twin techniques.Yuan et al. provided a systematic and comprehensive classification of various techniques to identify crack localization and pipe corrosion (Yuan et al., 2023).More recently Behari et al. conducted an analysis of the single and multiphase flow techniques for subsea and arctic conditions, where a summary of key findings from both experimental and field studies was presented however, there was still a lack of emphasis on mechanistic and CFD models as well as modern techniques such machine learning and digital twin based technologies (Behari et al., 2020).Studies for multiphase flow inside the tubing and with a leak for a wide range of hydrodynamic conditions such as viscosity, surface tension, and density, for a wide range of geometric conditions such as length and diameter of the tubing, for a wide range of operating conditions such as input volume fractions, input solid concentration, input size of the solids, input emulsion concentration, input pressure, and input temperature, and for a wide range of extremal environmental conditions such as flow direction, hydrostatic pressure are not available in the literature.We have previously conducted experiments, modeling and computational fluid dynamics simulations for single-phase flow (Rahman et al., 2007;Xiong et al., 2016), gas-liquid flow (Ahammad et al., 2018;Alfareq et al., 2020;Manikonda et al., 2020;M A Rahman et al., 2009;Mohammad A Rahman et al., 2009;Sleiti et al., 2021Sleiti et al., , 2020)), gas-liquid-solid flow (Alfareq et al., 2020;Barooah et al., 2021Barooah et al., , 2022aBarooah et al., , 2022b;;Khaled et al., 2021;Khan et al., 2021).A comprehensive study of varying leak sizes, orientations, locations, and numbers is also not available in the literature.Therefore, this work provides a review of the different studies conducted over the years on single-phase and multiphase leak detection using various techniques.Additionally, a primary concern with geological CO 2 storage is the potential for CO 2 leakage from faults, fractures, and abandoned well tubing.However, strategies for detecting such leaks are still in the early development stages.Therefore, this study also explores various techniques for CO 2 leak detection.
Leaks sometimes occur in the existing on-shore and off-shore (subsea) pipelines, as presented in Figs. 1 and 2, even when pipelines are designed with quality construction and appropriate regulations.In subsea conditions, leaks may happen due to (a) permafrost thaw settlement, (b) strudel scour, (c) ice gouging, and (d) upheaval buckling (Thodi et al., 2015).As presented in Fig. 3 (a) the leaks in subsea conditions are very challenging to deal with since it is rare events.A detection system should have higher sensitivity and reliability to prevent a catastrophic event.Leak prevention will also help the industry to reduce its carbon footprint.Qatar Petroleum launched an initiative in 2019 where CO 2 will be captured from natural gas facilities and sulfur recovery plants and reinjected to a subsurface sequestration site.Until 2019, 1.2 million tonnes of CO 2 were injected into a sequestration site.Qatar envisions reducing CO 2 emissions by 27% by 2030 as presented in Fig. 3(b).When any leak happens, it imposes a safety threat since it may spill hydrocarbon liquid into an ocean or natural gas can be dispersed to an environment.The economic impact of oil spills and natural gas dispersion through a leak can be huge.Failure to detect pipeline leaks promptly will have an adverse impact on life, the economy, the environment, and corporate reputation.Therefore, early leak detection, location, and size of a leak reliably and accurately are important aspects of efficient hydrocarbon transportation through a pipeline.To date, most efforts have been devoted to single-phase flow leak detection in benign environments, mostly single-phase onshore pipelines.Recently, the industry is considering using fiber optic cable (FOC) distributed sensing leak detection systems that can detect small chronic leaks.Fiber optic cable includes distributed temperature sensing (DTS) as well as distributed acoustic sensing (DAS) and distributed strain sensing (DSS) systems.However, finding the optimum location for the placement of fiber optic cable is a challenge.Additionally, the cost is huge in the case of an offshore leak and downhole conditions.Fig. 4 (a) provides the methodology for the Real-time transient monitoring (RTTM) approach which is an advanced approach for detecting leaks promptly and accurately.It is a continuous monitoring process with higher accuracy, and safety and provides early leak detection.However, the implementation and maintenance costs are higher, especially for longer pipelines, and false alarms and data interpretation can be challenging (Geiger et al., 2006(Geiger et al., , 2001;;Geiger and Vogt, 2014;Geiger, 2005).Therefore we propose using a combination of experimental, computational fluid dynamics (CFD), and mechanistic modeling for extending RTTM to data-driven modeling such as machine learning (eRTTM), (Hochgesang, 2018) (see Fig. 3).
New technologies are constantly emerging for leak detection, but there are still limitations to detecting a leak with a high degree of accuracy, sensitivity, reliability, and robustness while avoiding false alarms.The selected existing leak detection methods are presented in Fig. 4(b).It can be seen that the different leak detection techniques can be classified as externally based, internally based, and visual based.The next section provides a literature review on the different advancements in the state of the art for single and multiphase leak detection.

Advances in state-OF-the-art
A fundamental understanding of multiphase flow inside a pipeline is a challenging task.Any leak in single-phase multiphase flow and its dispersion to the surroundings is an added challenge.Failure to quickly detect a leak and its location can be catastrophic and can impose an adverse effect on the marine, economy, environment, and reputation of the operating company since it is very expensive and challenging to clean up oil spills and gas dispersion in the subsea environment (Zhang et al., 2013).Another challenge is to find the precise size and location of the leak (Palmer, 2000;Thodi et al., 2014Thodi et al., , 2015)).Fig. 5 shows a comparison of the different internal and external based leak detection techniques.Current leak detection systems based on internal pressure, negative pressure wave, temperature, and mass balance measurement (internal measurements) are good for identifying large pipeline leakage in benign environmental conditions.External measurements such as fiber optics acoustic, temperature, and strain sensors are also good in finding the location of small leaks (Abbaspour and Chapman, 2008;Doshmanziari et al., 2020;Harriott, 2011;Ling et al., 2015;van Bosselaar, 1967).Current leak detection technology is mainly based on single-phase flow and single detection techniques.However, there is a need to extend the current single-phase leak detection technology to multiphase flow leak detection using a number of different internal and external detection methods as real-time transient monitoring (RTTM (Bustnes et al., 2011;Chuka et al., 2016;Jahromi et al., 2019;Nicholas et al., 2017Nicholas et al., , 2015)),) and finally extending its reliability and robustness with artificial intelligence distinguishing an accurate "leak" and "no-leak" event.
The external and internal methods to detect leaks have pros and cons.External methods or physical inspection can result in accurate detection of the leak size and location but have a higher cost.Internal approaches detect gas leaks by solving the governing equations, thus leading to a lower cost, but with higher uncertainties (Martins and Seleghim, 2010).External methods include acoustic sensors, fiber optics, and vapor sensing.Acoustic sensors detect low-frequency noises that small leaks particularly produce (Khulief et al., 2012;Liang et al., 2013;Olivares, 2009;Stajanca et al., 2018).This method is liable to false alarms caused by other noises and would be expensive to use with extremely long pipelines.Fiber optic cables (FOC) are a good candidate for identifying pipeline leaks.Several studies have shown that fiber optics is potentially capable of detection and localization of very small leak rates down to 0.1% of the pipeline volume Fiber optic sensing mechanisms use sound, temperature, and strain to identify leak locations.By detecting the leak signatures fiber optics can detect leaks within seconds with a location accuracy of within 10 m.They are also Fig. 2. Fracture propagation of (i) gas (ii) liquid pipeline.(iii) pinhole leak in a pipeline due to corrosion (Mabily and Lehning, 2016;Yeom et al., 2014).(Mabily and Lehning, 2016;Yeom et al., 2014), (b) Different leak detection techniques (Adegboye et al., 2019;Zhang et al., 2013).
robust against false alarms due to pipeline damage (Baque, 2017;Siebenaler et al., 2017;Thompson et al., 2020;Velarde et al., 2020;Worsley et al., 2014).Fiber optics is better for onshore as compared to offshore applications.Erosion, deep-sea and arctic weather conditions, spacing, and support requirements limit the use of fiber optics in subsea or offshore applications.Environmental conditions interfere with DTS and ambient noise interferes with DAS.With vapor sensing, scans determine if vapors are emitted from a leak and the concentration of hydrogen is changed in an adjacent tube of the transportation line.This method detects the size and location of the leak, but since scans are only conducted periodically, it cannot find the leak instantly.Internal methods include statistical analysis, real-time transient modeling (RTTM (Nicholas et al., 2015)), volume balance, pressure drop, negative pressure wave, and so on.The statistical analysis looks at correlations between inlet and outlet flow and pressure.This method only works for limited cases and performs well in steady-state conditions (Hauge et al., 2009;Reddy et al., 2011;Syed et al., 2020;Wang and Carroll, 2007).RTTM measures irregular flow using mathematical models.This method works in transient conditions but must be monitored to reduce false alarms.Volume balance determines if there is a leak based on whether the overall mass remains the same (Stouffs and Giot, 1993).This method works well in transient conditions but cannot pinpoint the location of the leak.Pressure drops indicate a leak in a pipeline (bin Md Akib et al., 2011).This is useful for very small leaks, but it is not useful for determining where the leak occurred.Negative pressure wave sensors calculate where a leak occurred by comparing the arrival times of the waves to each sensor (Reynolds and Kam, 2019;Tian et al., 2012;Wan et al., 2011;Zhang et al., 2014).Zhang et al. (2014) used a dynamic pressure transmitter.The transmitter receives dynamic pressure signals at the inlet and outlet of a pipeline.Although they have performed different tests changing the operating conditions, their outcomes were not conclusive.Tian et al. (Zhang et al., 2014) used the negative pressure wave leak detection method.They also used a statistical method to analyze the data but their results were not conclusive due to the data quality.Several other researchers used transient pressure wave leak detection for crude oil pipelines, but the leak detection system experienced several false alarms (Abdulshaheed et al., 2017;Beushausen et al., 2004).

Multiphase flow leak detection
Pipeline leak detection system for single-phase flow is quite established as compared to multiphase flow systems.Multiphase flow systems involve the simultaneous movement of different phases with different properties, and variable flow conditions, making leak detection challenging due to the complex behavior of these fluids.Sensitivity to environmental factors, cost, data interpretation, and false alarms adds to the challenge.Additionally, the appearance of different flow patterns for changing input flow rates, volume fractions, and pressures makes leak detection challenging.Different flow regimes will impose different shear stresses, interfacial stresses, and flow fractions in a pipeline.Therefore, leak flow rates and leak hydrodynamics will be flow regime specific in the case of multiphase flow.Horizontal pipelines have stratified, bubbly, slug, and annular flow patterns (Ahn et al., 2019).Early empirical approaches to flow pattern delineation (e.g., Mandhane et al. (1974)) were followed by more reliable mechanistic modeling in later work.For example, Taitel and Dukler (Barnea et al., 1985) developed expressions for pattern transitions for horizontal/near-horizontal systems.Barnea et al. (1985) improved on that model by developing maps for inclined flow; further improvement was done by Barnea (1987) who offered a model for predicting flow pattern transition for the whole range of pipe inclinations.Kaya et al. (2001) offered further improvement in pattern prediction for deviated wells.Hasan and Kabir (Hasan et al., 2018) have summarized these developments in flow pattern delineation and offered their simplified approach to predicting flow patterns in all types of geometry in different inclinations.The power spectral density function (PSD), wavelet transform, Fast Fourier Transform, Hilbert-Huang transform, neural network approach, and others are among the most common signal analysis methods for the identification of multiphase flow regimes (De Fang et al., 2012;Elperin and Klochko, 2002;Fan et al., 2013;Park and Kim, 2003).Wavelet packet transformation has the potential to distinguish the different flow patterns.As mentioned earlier, nowadays industry is considering implementing the fiber optic technique for identifying both multiphase flow regimes and leaks, but the Fig. 5.The sensitivity, accuracy, and reliability of the different leak detection techniques (Adegboye et al., 2019;Zhang et al., 2013).Rahman et al. implementation of this technique is very expensive in subsea conditions (Ji et al., 2018a).Research on hybrid systems such as using multiple detection systems for multiphase flows is preliminary but promising (Lunger and Karami, 2019;Tavares et al., 2014).Table 1 provides a list of literature for pipeline multiphase flow leak detection experiments.Pressure sensors are the most commonly used tools for leak detection however, fiber optics and acoustic sensors can detect the smallest leaks.Acoustic sensors have certain limitations such as signal attenuation and distortion, extraction of signal and noise elimination can sometimes be difficult.In the case of optical fiber, it has a lower spatial resolution, higher cost, nondurable, and complex systems.The development of advanced algorithms for optimization of sensor structures can help to improve their limitations.
Pipeline leak detection in subsea arctic conditions imposes added challenges.Studies of complex pipeline networks in ultra-deep and subsea conditions are limited (Bai et al., 2017;Eisler, 2011;Eisler and Lanan, 2012;Kulkarni et al., 2012).If the subsea pipeline contains multiphase flow, the leak detection system is quite complex due to the acoustic noise coming from both the inside of a pipeline and externally to the pipeline, such as ice gouging, iceberg movement, and instrument malfunction.Bryce et al. (2002) implemented an arctic-based hydrocarbon vapor sensing leak detection system for a crude oil pipeline.The study was successful but the cost of construction, installation, operation, and maintenance along with the detection accuracy was a great challenge.Several industries adopted fiber optic systems to be deployed in subsea and arctic conditions, but again accuracy and cost are issues.

Pipeline leak detection using computational fluid dynamics (CFD)
The field pipeline is large in diameter with high temperature and pressure.Therefore, a numerical simulation can provide a better understating of field pipeline hydrodynamics, thermodynamics, and the consequences of pipeline leaks at different dimensions; reducing the number and cost of experiments (Cloete et al., 2009;de Vasconcellos Araújo et al., 2014;Zhu et al., 2014).Table 2 provides a list of literature on CFD and OLGA models for single and multiphase flow leak detection along with their key remarks.A 3-D turbulent flow model was implemented by Ben-Mansour et al. (2012) for a 10 cm diameter water delivery pipeline to detect a leakage (1 mm × 1 mm) using ANSYS Fluent.It was found that for steady-state simulations, small leaks do not greatly impact pressure variations.Moreover, small leaks can have a measurable impact on the magnitude and frequency of pressure signals for turbulent flow.Olivares (2009) looked at the impact of atmospheric pressure and temperature on leak noise in a district water heating pipeline.De Vasconcellos Araujo et al. (de Vasconcellos Araújo et al., 2014) used ANSYS CFX for a leak detection study.They found that leaks at a closer distance have higher pressure drop they do not have any effect on the input pressure.Zhu et al. ( 2014) created a computational model to predict oil leakage from a broken submarine pipeline.They observed that for a larger density of oil, the leak is smaller.The oil spill process consists of a rising and drifting process.Cloete et al. (2009) performed a 2-D transient simulation using Fluent for a subsea gas pipeline.The outcome of the model was forwarded to the ANSYS Static Structural model to explore the fluid/structure interaction (FSI).Jia (Jia et al., 2012) looked into slug-fluid induced vibration (FIV) for the pipeline, jumper, and riser components.Pontaza and Menon (2011) and Elyyan et al. (2014) described a screening technique for flow-induced vibration (FIV) caused by internal multiphase flow to measure the fatigue life of the jumper.The multiphase flow simulator OLGA was also used by a number of researchers for leak detection (Afebu et al., 2015;Syed et al., 2020) The adaptive Luenberger-type observer was used in OLGA for two and three-phase flow leak detection (Afebu et al., 2015).The OLGA observer was integrated with Matlab, which assisted in providing different input boundary conditions in OLGA.CO 2 transport pipeline in the presence of excess water also tends to form hydrates at low temperatures (Peters et al., 2012).Multiphase flow assurance in an offshore pipeline, as well as its effect on offshore pipeline integrity, are still in the development phase (Guo et al., 2005;Kak et al., 2017;Leporcher et al., 1998;Lunde et al., 2013;Sloan Jr and Koh, 2007).The study of the effects of flow patterns on leak detection in multiphase flow conditions is still in the early stages of understanding.Extensive CFD simulations are therefore needed to generate well-supported models.

Pipeline leak detection using mechanistic modeling
There are limited studies available for single-phase flow pipeline leak detection using mechanistic modeling (Gajbhiye and Kam, 2008;Kam, 2010a;Thiberville et al., 2018).A few studies are also available for the multiphase flow pipeline leak detection (Thiberville et al., 2019).Pipeline leak detection for long pipelines using sensors/instruments is difficult.Observing differences in up and downstream rates and/or pressure is one of the potential techniques for leak detection.Fluid flowing from one end of a transportation line to the other generates a pressure change, ΔP.For horizontal systems, ΔP is dependent on, among other variables, the square of the flow rate.A leak that allows fluid to flow out of the system causes a change in line ΔP.Table 3 provides a list of literature on mechanistic correlations for leak detection along with their key remarks.Most of the studies have used changes in flow rate and pressure for the detection of leak size and location.Several studies on pipeline leak detection methods have been proposed in the literature for single-phase fluid flow based on this change in line ΔP (Ferrante et al., 2014;Liu et al., 2019;Rougier, 2005).A similar transient modeling approach using changes in ΔP was applied by Hasan et al. (Hasan et al., 1998;Kabir and Hasan, 1998) to detect flowline blockage.A slightly different approach, single and multi-rate tests, was used by Rui et al. (2017) for gas pipelines.However, as mentioned earlier, these investigations are limited to single-phase fluid flow.Very few researchers applied this modeling approach to pipelines carrying two-phase fluid.Kam (2010b) investigated the influence of leak sizes and their location for a multiphase flowing line.Similarly, Figueiredo et al. (2017) studied leakage for two-phase flow in nearly horizontal systems and found that the approach used for single-phase systems is also applicable for multiphase flowing systems.However, these studies used empirical multiphase flow correlations, which are unreliable under many conditions, especially in subsea situations.

Leak detection in subsurface sequestration sites
There are several industrial-scale active CO 2 storage projects around the world including in Qatar (Buscheck et al., 2012;Cihan et al., 2011;Kermani and Daguerre, 2010;Spinelli et al., 2014).However, there are uncertainties regarding the risks of this technology playing an important role in managing anthropogenic CO 2 emissions (Lucci et al., 2011;Meltzer et al., 2014;Rangriz Shokri et al., 2019;Sun et al., 2016;Veltin and Belfroid, 2012).The main risk associated with geological CO 2 storage is the CO 2 leakage from the faults and fractures and abandoned well tubing (Imrie et al., 2019;Leonenko and Keith, 2008).As presented in Fig. 6, the wellbore consists of several concentric tubular columns where the sensors can be mounted to monitor the pressure, temperature, and phase change signals.There are several techniques for leak detection in a CO 2 sequestration project, such as well pressure monitoring,

Table 2
List of literature using CFD models for leak detection.employing a model-based approach using seismic-pressure data, predictive modeling of injected CO 2 plume in the subsurface, and detection of near-surface gas sampling (Al-Hussain et al., 2015;Liu et al., 2018;Mohamed et al., 2012;Sun et al., 2015;Venna et al., 2018).However, in practice, two technologies are common to monitor the leak in deep subsurface oil and gas wells: downhole and surface technologies (Al-Hussain et al., 2015).Table 4 provides a list of literature for multiphase flow leak detection in CO 2 injection wells using various downhole and surface technologies along with their key remarks.The surface technology includes time-lapse seismic surveys, while the downhole technology includes tools such as fluid samplers, micro seismic sensors, downhole pressure and temperature gauges, and distributed fiber optic sensing (Al-Hussain et al., 2015).Several statistical and machine learning approaches exist for downhole leak detection (Sinha et al., 2020;Yang et al., 2019).The quantification of pressure anomalies resulting from leakage such as in injection and abandoned wells can be performed using a large number of analytical and numerical modeling approaches (see Fig. 7).

Leak detection using machine learning and digital twin
Accurate physical simulations typically require massive resolution and associated computational requirements.Simpler models that target the key aspects of a modeling problem must be chosen for practical purposes, such as when real-time calculations are required to give instantaneous predictions or system analytics.Simplification must be performed carefully to not trade off too much accuracy so that the results no longer model the target physical system well.One recently emerged method to simplify compute-intensive physical modeling problems involves using machine learning (Afebu et al., 2015;Alves Coelho et al., 2020;Venna et al., 2018).As presented in 7 (a) and 7 (b), artificial neural network (ANN) and machine learning (ML) approaches were used for single-phase leak detection.The basic idea is that a high-resolution physical measurement or some concrete simulations of a system can represent a set of training data for the complex function represented by the simulation problem.From this data, machine learning can model the underlying complex function that models the physical phenomenon.However, ML approaches can have certain drawbacks such as acquiring actual operational data is difficult, sometimes ML models are not suitable for practical applications and it has higher upfront cost and is time consuming.To counter these drawbacks the emerging machine learning model can be as a computationally more efficient prediction or analysis mechanism to facilitate practical usage of the Digital Twin (DT).Digital twins are a paradigm that combines multi-physics modeling with data-driven analytics and uses them to simulate a twin's life.Piltan and Kim (2021) proposed a combination of fuzzy digital twin, support vector machine (SVM), and backstepping (BS) for identifying the location and size of cracks in pipelines.This model was able to predict cracks in the pipeline with 25% improved accuracy.Chen et al. (2022) provided an initial vision of the digital twin of the subsea pipelines with coverage of current applications.However, the collection, interpretation, sharing of data, and cyber security remain some of the primary challenges.Wang et al. (2023) proposed a combination of a digital twin model and support vector machine (SVM) of a pipeline based on pressure signals generated and were able to predict pipeline gas leaks with high accuracy.An example of DT architecture for CO 2 pipeline transport systems that can be extended to oil and gas pipelines was proposed recently by Sleiti et al. (2022).Once fully implemented, this DT can be used for pipeline safety monitoring including leak detection, condition-based maintenance, life cycle prediction, cost reduction, and autonomous operation of pipeline transport systems.In addition, the DT is intended to assess the remaining life for various parts, and hence for planning a service interruption or prediction of a fault based on historical data and real-time measurements.
In multiphase fluid flow in pipes, underlying models of particular complex flow regimes either do not exist or require direct numerical simulations (DNS) of Navier-Stokes equations that are far beyond current computational capabilities.Simplified, alternate approaches exist Fig. 6.Subsurface leak monitoring for a CO 2 injection well (De Jong et al., 2019).

Table 4
List of literature for CO 2 injection well.

Fluids used
Model used Leak in terms of Focus Key remarks Reference Liquid-liquid, Gasgas, Gas-liquid

Support vector machine
Temperature and spectral noise logging.

Downhole leak
The developed model was validated with field data with 91% accuracy.
A combination of spectrum estimation, principal component analysis, and Support vector machine outperformed the Support vector machine.

Downhole leak
The developed approach has higher accuracy in determining any corrosion or leak depth as compared to the rigless approach.

Al-Hussain et al. (2015) Gas-liquid Integrated approach
Temperature and noise logs.

Wellbore
The temperature log alone is not sufficient, while an integration of temperature and spectral noise logging techniques is effective in locating and characterizing leak sources.

Deep subsurface storage formation
The developed approach was able to predict the existence and location of leaks.
The approach is applicable if the average reservoir condition during the testing period remains stable.(e.g.turbulence closures such as Reynolds-averaged Navier-Stokes and Large Eddy simulation) however, for each regime, they must be validated against real fluids (Brunton et al., 2020a).This ultimately requires high-quality experimental data.
An alternate approach that has not been well-explored is based on a data-driven statistical approach.However, similar success in application to fluid flow modeling has not yet been achieved, as such data is not widely available and must be collected with highly specialized experimental flow loops (Brunton et al., 2020b).

Recommendations
A comprehensive review of the different leak detection techniques, their applications, and their limitations suggests that there is a need to utilize new aspects of fluid flow, in particular, tailored toward the modeling of leaks.There is a need to explore how both experimental data and subsequently produced CFD simulation data can be used to drive a statistical approach based on modern deep learning techniques.Both supervised and unsupervised techniques can be applied to this task in order to create effective models to detect and predict pipeline system leaks.Historical data measured from sensors in a working pipeline system can be incorporated into these predictions, using anomaly/irregularity as another factor in the prediction of leaks in a given system (Candelieri et al., 2014;Moubayed et al., 2021).The existing work suggests that escaping flow rate, pressure, and pressure gradient generally have a sharp signature variation near the leak orifice location (Jujuly et al., 2016).If accurate, the experimental and CFD data can provide insight into the signatures of a leak to determine its location, size, and severity (Candelieri et al., 2014).However, weighing the various signatures and flow features and how they map to the desired leak detection outputs analytically becomes overwhelming by sheer human analysis as the number of simulations and parameter space explored becomes large enough.It is recommended that machine learning methods be employed in order to automate the creation of leak detection and prediction functions.Both simulation (CFD) and experimental data over a range of flow conditions and pipeline geometries can be generated and used to supervise a learning process to train a leak detection and localization function (Alves Coelho et al., 2020;Oliveira et al., 2018;Sfar Zaoui et al., 2020;Van der Walt et al., 2018).Supervised learning methods are known to require large amounts of data (Arsene et al., 2012;Chen et al., 2017).Given that certain flow signatures such as flow rate change, acoustic anomaly, temperature gradient, external ground change, and pressure variation are known to correlate with leaks based on existing studies, it may make sense to provide these features directly as input to the machine learning system for training to decrease the complexity of the function learning process (Banjara et al., 2020;Bhowmik, 2019).It may also be possible to use unsupervised learning techniques (e.g.autoencoders or similarly-inspired networks) to extract interesting features in the training dataset that can serve as automatically learned signatures that can be re-used as features.This approach has been taken in some existing fluid mechanics work.Ideally, the resulting leak detection module can serve as a universal leak detection method across a large range of flow and pipeline parameters.
An alternate and perhaps complementary perspective towards solving the problem of leak detection will be to look at characterizing hazards such as leaks as novel or anomalous events that deviate from normal observed behavior.This approach is described in previous pipeline monitoring work (Rahmes et al., 2015) but does not appear to have been explored and implemented fully in the context of multiphase fluid flow where "normal" flow behavior may vary widely as various flow regimes may be observed as "normal" in a particular installation.This strategy may likely be combined with the supervised learning approach described above potentially as two possible strong signaling mechanisms used to gauge and detect leaks.

Conclusion
Leaks of natural gas and oil can be extremely detrimental to the economy.If pipeline leaks are not promptly discovered, the economy, the environment, human health, and corporate reputation will all suffer.Therefore, this paper provides a comprehensive review of the different leak detection studies conducted over the years.Our findings can be summarized as follows.
• A comprehensive review was conducted where various technologies such as analytical, mechanistic, CFD, and machine learning models were studied, and their key results and limitations were summarized.Furthermore, it was also revealed that there is still a need to improve our existing knowledge of multiphase leak detection.It is conceivable that the number of publications per year will increase, and that leak detection techniques based on software, machine learning, acoustic, and optical fiber will continue to be in demand in the coming years.• Although the hardware-based techniques have higher accuracy and sensitivity in locating leaks, however a higher cost, nondurable and complex systems, signal extraction, and denoising reduces their feasibility.• Software-based approach has higher promise due to their lower cost and maintenance, however their accuracy in detecting leak size and location is far from ideal currently.• Detection of smallest leaks and their location still remains one of the most difficult challenges.A combination of hardware, software, and AI techniques is the way forward to tackle this problem.• A set of guidelines is provided that states how new emerging technologies like machine learning and digital twin can improve our current understanding of multiphase leak detection thereby cutting down costs by ensuring regular supply, improving long-term production stability, and reducing the final costs of petroleum products, thereby saving millions of dollars.(Afebu et al., 2015).(b) Acoustic leak detection using ML approach (da Cruz et al., 2020).

Table 1
Selected pipeline leak detection experiments in the literature.
Higher accuracy was obtained in predicting leak locations than leak sizes.A large data set is required for accurate prediction of leak sizes.Leak prediction improves for larger leak sizes, more leak distance from inlet, and lower leak numbers.

Table 3
List of literature using mechanistic correlation for leak detection.